Sinter basicity directly influences its strength, reducibility, and soft-melting properties by regulating mineral composition and structure, acting as a critical control index to ensure stable blast furnace operation and improve smelting efficiency. Predicting sintered ore basicity is an important way to overcome the inherent lag of industrial processes and the delays of traditional testing methods. To address the shortcomings of existing deep-learning prediction models, namely structural rigidification and overfitting during training, a regularised self-organising gated recurrent unit (RSO-GRU) model is proposed for sintered ore basicity prediction in this study. The model incorporates an adaptive regularisation mechanism based on weight dispersion and the rate of change of prediction error, dynamically adjusting regularisation strength according to the network's training state, thereby mitigating the structural rigidity of conventional deep-learning architectures. A structure self-organising growth strategy driven by neuron-sensitivity analysis is employed to match network size to task complexity, and the algorithm's convergence is proven via Lyapunov stability theory. Experimental results demonstrate that the RSO-GRU model achieves an average absolute error (MAE) of 0.0158, a root-mean-square error (RMSE) of 0.0201, and an average absolute percentage error (MAPE) of 0.8126%, representing reductions of 16.4, 15.2 and 16.6%, respectively, compared with a standard GRU model, while its coefficient of determination ( R 2 ) rises by 5.8% to 0.8505, confirming the significant optimisation effect of the proposed regularised self-organising strategy. Comparative studies against SVR, RF, BP and LSTM algorithms show that RSO-GRU outperforms all benchmarks across evaluation metrics, meeting the precision requirements for industrial applications.
Xie et al. (Fri,) studied this question.